| The rapid identification and accurate positioning of navigation feature targets are the key technologies of visual navigation.For the trellis-type kiwi fruit orchard,the kiwi fruit trunk is used as the detection object,and the kiwi fruit orchard’s inter-row navigation feature target sample library is constructed through image preprocessing and detection object labeling;the three-dimensional geometric conversion relationship of target positioning is established through the internal parameters of the camera;Scale detection ideas,increase the feature extraction results of shallow neural networks,set up multi-scale feature map RPN network regression,build an improved Faster R-CNN network structure,achieve accurate recognition of small-sized trunks at the far end of the tree row,and expand the single The number of side tree row trunk recognition;using the characteristics of straight line planting,proposes an alternative method for the root point of the target detection frame,which maintains the positioning accuracy and simplifies the positioning algorithm;adopts the cubic spline interpolation method of the root point coordinates to obtain the tree by fitting The line line is intercepted by the equidistant points of the tree line on both sides to determine the navigation base point between the lines,and then the least square method is used to fit the line navigation line.The main research contents and results are as follows:(1)Navigation camera calibration and target sample database construction.Using Matlab software,combined with Zhang Zhengyou’s camera calibration method,determine the internal and external parameters of the camera;according to the straight-line planting characteristics of kiwi orchard trees,tree trunks or cement pillars are used as navigation feature targets,and the navigation distance requirements between the rows and the feature target detection scale requirements are selected.The image size is 1500×900 pixels.The sample labeling rules are formulated according to the shape characteristics of the navigation feature target,and 1600 tree trunks and 1200 cement pillar images in the training set are marked,and 10162 tree trunk samples and 7440 cement pillar samples are obtained.(2)Based on Faster R-CNN small target detection improved algorithm construction.The difference between the low-level features and the high-level features of the image is analyzed from the perspective of visualization.For small target detection,a combination of the low-level features extracted by the shallow network and the high-level features extracted by the deep network is proposed to construct the multi-scale feature regression idea of the candidate area.Improve the VGG-16 structure and RPN network structure of the Faster R-CNN backbone network,divide VGG-16 into 5 stages of convolutional layers,remove the fully connected layer and softmax layer after the 5th stage,and extract the convolutional layers of each stage The multi-scale feature map of,is input into the corresponding RPN network for target frame regression.The test results show that the improved algorithm for small target detection based on Faster R-CNN has a correct recognition rate of 83.6% for small target tree trunks with a pixel height of 100-200,which is approximately higher than that of Faster R-CNN,YOLO v3 and SSD.17%,27% and30% indicate that the algorithm effectively improves the detection effect of small targets.The correct recognition rates of trunks under mulching,soil and weed environments were90.05%,89.55%,and 91.2%,respectively,which were about 10%,18% and 35% higher than Faster R-CNN,YOLO v3 and SSD,respectively.The algorithm environment is adaptable.(3)Propose a target location method for kiwifruit trunk root replacement.Taking the midpoint of the bottom edge of the target detection frame as the base point of trunk positioning,simplifies the positioning of the actual trunk root point;taking the actual trunk root point as the origin,the distance deviation of the obtained trunk positioning base point is analyzed to verify the accuracy range of the root point substitution method.The test results show that the average value of the horizontal and vertical deviation is 8.9 pixels,the pixel deviation accounts for 1.2%,the linear distance between the two coordinates is 12.6pixels,and the average actual distance error is 0.096 m.(4)Tree line generation and navigation line generation and result analysis.First of all,in view of the irregular arrangement of the trunk positioning base points obtained by the root point substitution method,the advantages of the cubic spline interpolation method,such as local controllability and continuity,are used to achieve the smooth processing of the unilateral trunk positioning base points and reduce the fitting tree line.The set deviation between the line and the actual tree line reduces the set deviation between the fitted tree line and the actual tree line;the method of equidistant interception of the fitted tree line on both sides is used to calculate the fit of the center line of the navigation between rows Base point,use least square method to generate navigation line between rows of orchard.The experiment showed that the maximum horizontal deviation of the midpoint of the navigation line in the three environments of plastic film,soil and weeds were 15.3 pixels,16.7 pixels,and 19 pixels,respectively,and the relative error of the average horizontal deviation was 1.2%.The average lateral deviation is 7.1 pixels,7.6 pixels,and 8.2 pixels,respectively.The maximum heading angle deviations are 8.1°,8.6°,and 8.7°,and the average heading angle deviations are 4.2°,4.5°,and 4.6°. |
